Weakly Supervised Breast Tumor Detection and Segmentation via Grad-CAM++ Guided Mask R-CNN

The analysis of breast cancer histopathology is challenged by complex tissue morphology and the scarcity of detailed annotations. This work introduces an end-to-end computational framework to address these limitations by enabling precise tumor segmentation and detection without pixel-level ground truth. Our approach first enhances image quality through adaptive techniques and then leverages a high-accuracy classification model to establish context. The core of our method employs Grad-CAM + + to generate class-discriminative localization maps, which are converted into initial pseudo-masks. These masks are then substantially refined by a loss-modified Mask R-CNN, trained under weak supervision. Evaluated on the BreaKHis dataset, the framework achieves a Dice score of 0.872 and an mAP@0.5 of 0.837, demonstrating superior performance and robustness across image magnifications. This strategy significantly lessens the dependency on expert annotations, presenting a viable path toward integrating automated analysis into digital pathology workflows. The implementation code for this study is publicly available at our GitHub repository: https://github.com/Maisamilens/breast-cancer-segmentation.